Toward machine-assisted tuning avoiding the underestimation of uncertainty in climate change projections

Documenting the uncertainty of climate change projections is a fundamental objective of the inter-comparison exercises organized to feed into the Intergovernmental Panel on Climate Change (IPCC) reports. Usually, each modeling center contributes to these exercises with one or two configurations of its climate model, corresponding to a particular choice of "free parameter" values, resulting from a long and often tedious "model tuning" phase. How much uncertainty is omitted by this selection and how might readers of IPCC reports and users of climate projections be misled by its omission? We show here how recent machine learning approaches can transform the way climate model tuning is approached, opening the way to a simultaneous acceleration of model improvement and parametric uncertainty quantification. We show how an automatic selection of model configurations defined by different values of free parameters can produce different "warming worlds," all consistent with present-day observations of the climate system.